BUFFALO, N.Y. -- A suspected crooked insider at a New York software company sells consumer-credit reports to identity thieves, at roughly $30 a pop, in a high-tech scam that prosecutors say victimizes thousands.

An unemployed British computer administrator fights extradition to face federal charges in Virginia and New Jersey that he hacked into 92 separate U.S. military and government networks, often getting past easy-to-guess passwords to download sensitive data.

These and other recent data intrusions, whose authors are typically intent on theft, sabotage or cyberterrorism, have given rise to a promising profiling-and-reasoning strategy aimed at preventing online break-ins as they happen.

Just as authorities use profiling to guard against criminals at ports and borders, researchers at the State University of New York at Buffalo are developing software that can generate highly personalized profiles of network users by analyzing the sequences of commands entered at each computer terminal.

The system -- a prototype is likely to be ready for intensive testing this summer -- could provide a high-grade layer of protection for military installations and government agencies as well as banking or other commercial networks that require especially tight monitoring.

The "user-level anomaly detection" software draws up regularly updated profiles by closely tracking over time how each person performs an array of routine tasks, such as opening files, sending e-mail or searching archives.

Designed to tell if someone has strayed into an unauthorized zone or is masquerading as an employee using a stolen password, the program keeps watch for even subtle deviations in behavior.

Alerted to anomalies, network administrators then begin monitoring more aggressively to assess whether pilferage is in progress.

"The ultimate goal is to detect intrusions or violations occurring on the fly," said chief researcher Shambhu Upadhyaya, a SUNY Buffalo computer science professor. "There are systems that try to do this in real time but the problem is it results in too many false alarms."

Keeping false alarms to a manageable minimum is key, but extremely difficult to achieve, said Bruce Schneier, a network security and cryptography expert and author of Secrets & Lies, Digital Security in a Networked World.

"These systems live and die on false alarms," said Schneier. "You see this problem in facial recognition, trying to catch terrorists by recognizing faces in airports. All those trials failed miserably."

The Buffalo school is one of 36 research and teaching centers designated by the National Security Agency since 1998 to help safeguard America's information technology systems.

Aided by doctoral student Ramkumar Chinchani and Kevin Kwiat of the Air Force Research Laboratory in Rome, N.Y., Upadhyaya in 1999 began examining whether monitoring simple user commands instead of network traffic might produce faster, more effective monitoring.

Some computer-security products that feature user-profiling seek out deviations on the basis of huge amounts of data flowing through entire networks. They're typically 60 percent to 80 percent reliable, whereas simulation tests indicated the new software would be up to 94 percent reliable, Upadhyaya said.

The software borrows from risk-analysis economic models, drawing on dynamic reasoning and engineering methodologies. And even if it proves successful, the software would be just one tool in a computer-security arena that requires multi-layered defenses, Upadhyaya said.

"Hackers are a step ahead of you always," he explained, noting that the military "is especially worried about the insider who's been there a long time and learned all the loopholes."

Mike Kurdziel, an information security specialist at Harris Corp., which makes tactical military radios, thinks Upadhyaya has "constrained the problem" by installing various thresholds to curtail false alarms.

"Other intrusion techniques require something like looking at audit logs after the damage has already occurred," Kurdziel said. "The advantages offered by this approach is an intruder with malicious intent can be identified very early and a system operator can contain the damage, repair it in real time and shut out the intruder."